Characterization of a Tree Mapping Algorithm for Tree-to-Tree Transducer Induction

Abstract

We characterize a tree mapping search space in terms of the tree fragment depth and number of variables, which are parameters of the resulting tree transducer grammar. We show how such characterization explains the trade-off between computational complexity and tree transducer expressivity. We evaluate our induced tree transducers on a Question-Answering task, quantifying accuracy and average tree mapping time as a function of our parameterization.